Closed DuHao55 closed 2 months ago
You can just load the pictures from your dataset to replace dataset = datasets.ImageFolder(os.path.join(data_path, 'val'), transform=transform)
in class EffectiveReceiptiveField
in utils.py
. I think the conclusion would not be that different from classification.
But if you really want to derive into the task segmentation from the perspective of ERF, I recommend you to use the checkpoint tuned in segmentation and see if it differs from the same model trained on classification.
Thanks,I still don't understand what you mean. I am using it for segmentation task and the dataset format is as follows: ├── dataset │ ├──images │ │ ├── train │ │ └── val │ ├── labels │ │ ├── train │ │ └── val How can I change the EffectiveReceiptiveField in utils.py?
So how you get the dataset instance for segmentation?
After you get that, you can replace the line
with your own built dataset. and that is all you need to do if you just want to change the dataset.
Yes, your comment is correct. I noticed that ImageFolder doesn't work for datasets similar to VOC format (ImageFolder needs some subfolders) and I don't know how to load datasets similar to the above format. May I ask how I can modify the way line 328 datasets are loaded? ├── dataset │ ├──images │ │ ├── train │ │ └── val │ ├── labels │ │ ├── train │ │ └── val
May I know how to use erf.py to draw erf about segmentation task? I noticed that my dataset format(like VOC format) is not quite the same as the dataset format for the classification task.